1. University of Washington, Autumn 2018
Call Data Records
Lecture 28: CSE 490c
12/3/18
University of Washington, Autumn 2018

2. University of Washington, Autumn 2018
Topics
Data Science for Development
AI for Social Good
Today
Call Data Records
12/3/18
University of Washington, Autumn 2018

3. University of Washington, Autumn 2018
Announcements
Homework 7 Due Tonight
Programming Assignment 4 Due December 11
You have received the teaching evaluation link. Complete evaluations by December 9
A high response rate is very important for meaningful results. We will send reminder s to non-responders during the evaluation period. In addition, studies show that instructor involvement can increase response rates as much as 15-20%.
12/3/18
University of Washington, Autumn 2018

4. University of Washington, Autumn 2018
Telco Information
Call Data Records (Call Detail Records)
Meta data on individual calls
Cell Tower Logs
Information of handset connections with towers
Cell Tower Locations
12/3/18
University of Washington, Autumn 2018

5. Access to cell phone data
Proprietary to Telco
Provide competitive advantage
Possibly for marketing or data services
Linkage with mobile money
Subject to government privacy regulations
Possible access to aggregated data
12/3/18
University of Washington, Autumn 2018

6. University of Washington, Autumn 2018
Reading for this week
An Investigation of Phone Upgrades in Remote Community Cellular Networks, Kushal Shah et al., ICTD 2017
12/3/18
University of Washington, Autumn 2018

7. University of Washington, Autumn 2018
Call Data Records
Meta data associated with calls
Source number
Destination number
Source Tower (ID)
Destination Tower (ID) [might be missing]
Time
Duration
Status
12/3/18
University of Washington, Autumn 2018

8. University of Washington, Autumn 2018
Types of studies
How people use technology
Populations studies (where people are)
Event studies (what happens when)
Epidemiology studies
Economic studies
Distinction between Aggregate Studies and deriving information about individuals
12/3/18
University of Washington, Autumn 2018

9. University of Washington, Autumn 2018
Working with CDRs
Preprocess data for higher level structure
Align data with other sources
Tower data
Economic / Population Data
Compute home location
Determine movement patterns
12/3/18
University of Washington, Autumn 2018

10. University of Washington, Autumn 2018
Call Graph
Directed or undirected graph on calls
Measurement of call volume
Detection of high in- degree and out-degree nodes
Identification of social network
12/3/18
University of Washington, Autumn 2018

11. University of Washington, Autumn 2018
Call Graph Analysis
Global analysis versus individual analysis
Is the goal to understand aggregate properties or individual properties
Feature identification of individual’s calls
Number of calls
Length of calls
Missed calls
Incoming vs out going
Time of day
Neighborhood
Frequent caller neighborhood
Nodes of distance two
12/3/18
University of Washington, Autumn 2018

12. Social Network Identification I
How do you identify a callers Social Network from a call graph?
Start with the Induced Subgraph on the Neighborhood of the individual
12/3/18
University of Washington, Autumn 2018

13. Social Network Identification II
Identify highly connected groups of vertices in Neighborhood Graph
Finding a maximum Clique is NP-Complete
Heuristics
Maximum degree subgraph
Maximum density subgraph
12/3/18
University of Washington, Autumn 2018

14. Degree-K Subgraph problem
While there is a vertex of degree less than K
Delete all vertices of degree less than K
12/3/18
University of Washington, Autumn 2018

15. Maximum Density Subgraph problem
Find an induced subgraph S that maximizes ration Edge(S)/Vertices(S)
Polynomial time algorithms using Network Flow techniques
Related to degree K subgraph problem
12/3/18
University of Washington, Autumn 2018

16. How good a predictor is the call graph of an individual's income?
12/3/18
University of Washington, Autumn 2018

17. University of Washington, Autumn 2018
Phone use studies
Given demographic information, study phone use behavior
Call volume, call timings (time of day, date), number of contacts
Studies of Sim Card Churn
Inference of demographics from behavior
Mehrotra et al. (2012), Differences in Phone Use Between Men and Women: Quantitative Evidence from Rwanda.
12/3/18
University of Washington, Autumn 2018

18. University of Washington, Autumn 2018
Migration Studies
Movement of people is an area of significant study
Lack of census data make this hard to study
Short term migration
What are the patterns
Is it possible to distinguish between shorter term and permanent migration
Forced migration and droughts
Match to climate data
Question on local versus long distance migration
Technical issues in definitions of movement
12/3/18
University of Washington, Autumn 2018

19. University of Washington, Autumn 2018
Blumenstock et al. (2015), Predicting poverty and wealth from mobile phone metadata
Economic Studies
Predict economic status at a local (e.g. District) level
Household surveys are expensive. Idea is to use Cell Phone Data to expand surveys
Correlate household surveys with CDR
Compute wide range of properties of CDR
Construct machine learning model to predict household assets (from survey data)
Apply to all call records in the data set
12/3/18
University of Washington, Autumn 2018

20. University of Washington, Autumn 2018
Epidemiology
Correlating human movement data and disease frequency
Substantial work on Malaria and Call Data Records
Key use case is malaria elimination
Understanding if cases are local infections or from other regions
Understand movement from high incidence to low incidence areas
Technical modelling work that combines migration and economic studies
12/3/18
University of Washington, Autumn 2018

21. University of Washington, Autumn 2018
Event studies
Look at impact of events in data sets
High volume of calls related to disasters, elections, holidays
Spike in call volumes has been observed associated with earth quakes
Significant interest in call data records and disaster response
Technical issues related to infrastructure and economic displacement
12/3/18
University of Washington, Autumn 2018

22. Additional challenges on CDR Analysis
Countries often have multiple Telcos
Getting data from all Telcos is even harder
Is data from one Telco sufficient?
Increasing use of other media for communication
Encrypted messaging Apps such as WhatsApp
Analysis of Social Media company data should be even more restricted than CDR
Aleksandr Kogan
12/3/18
University of Washington, Autumn 2018